Elephant random walk with attributed steps and extractions of random sizes

Abstract: We study a model of market economics wherein the $(n+1)$-st customer, for each $n\geqslant N$, with $N$ being a prespecified positive integer, draws a sample of (random) size $K_{n}$, either with replacement or without, from the customers of the past. Each sampled customer is queried as to which of the two products, A and B, available in the oligopolistic market, they chose, and whether they are satisfied or not with their choice. The $(n+1)$-st customer now employs a stochastic rule, based on the information collected from the sampled customers, to decide which of the two products to buy. The probability that a customer is satisfied with the product they have purchased equals $q_{1}$ when the product is A, and $q_{2}$ when it is B, independent of all else. The resulting stochastic process may be represented as a variant of the celebrated elephant random walk, with the relative performance (in terms of sale) of A with respect to B, up to and including the $n$-th sale, captured by the position $S_{n}$ of the walker at time $n$. We study the almost sure convergence of $S_{n}/n$, as well as the convergence in distribution of suitably scaled versions of $S_{n}$ (where the scaling depends on the regime we are in).

• The paper presents a novel ERW extension integrating attributed steps and randomized memory extractions to capture customer choice dynamics.
• It employs stochastic approximation and ODE methods to establish almost sure convergence and detailed fluctuation regimes under varying reinforcement conditions.
• The findings offer explicit criteria for unique limit proportions and Gaussian fluctuation behaviors, with practical applications in market modeling and adaptive systems.

Elephant Random Walks with Attributed Steps and Random-Size Memory: Model and Asymptotic Analysis

Introduction and Motivation

This paper ("Elephant random walk with attributed steps and extractions of random sizes" (2604.17302)) extends the theory of memory-driven random walks, particularly focusing on the so-called Elephant Random Walk (ERW), by introducing two innovations: (i) step attributes reflecting binary customer satisfaction, and (ii) extraction of memories (sampled past steps) of random cardinality. The context is an oligopolistic market, modeling the dynamics of customer choice between two rival products, incorporating both memory of peer choices and response feedback.

The process generalizes both classical ERWs and generalized urn models, subsuming reinforcement mechanisms with sample sizes determined by arbitrary random laws. The primary focus is on rigorous characterization of strong and weak convergence of normalized process statistics, including scaling limits and the elucidation of deterministic ODEs governing the macroscopic evolution. The authors employ stochastic approximation techniques to derive almost sure (a.s.) convergence, central limit theorems, and fluctuation characterizations, unifying the memory-reinforced random walk paradigm with path-dependent urn models with arbitrary reinforcement.

Model and Formalism

Stochastic Process with Attributed Memory

At each time step $n+1$, a new agent (customer) enters, samples a (random) subset, $K_n$, of previous $n$ steps (customers), and observes both (a) their chosen product ($A$ or $B$) and (b) a binary satisfaction attribute. The agent's decision to choose $A$ or $B$ is governed stochastically by an arbitrary smooth reinforcement function $F$, which aggregates the observed proportions in the sampled set. The satisfaction probability is $q_1$ for $A$ and $K_n$0 for $K_n$1, independent of history.

Sampling can be with or without replacement, and the law $K_n$2 of $K_n$3 may be either fixed (iid) or time-varying, supported up to $K_n$4 in the without-replacement case.

Mapping to Random Walk and Urn Framework

The process is encoded as a random walk (RW): $K_n$5 if product $K_n$6 is chosen, $K_n$7 otherwise; $K_n$8 tracks net $K_n$9 sales. The process can also be construed as a generalized balanced four-color urn process, where ball colors track customer/product/satisfaction classes. At each epoch, a sample of $n$0 balls is drawn, colors are recorded, and a new ball is added whose color is assigned according to a probability function based on sample color counts.

This formulation gives rise to a complex, non-Markovian, path-dependent stochastic process with generalized reinforcement and random memory length.

Main Theoretical Results

Almost Sure (Strong) Convergence

General Case

• In both sampling schemes (with/without replacement), and for both fixed and time-varying laws $n$1, the proportions $n$2 converge almost surely to a deterministic limit set, which can be further refined under contraction-type conditions.
• The limiting behavior is governed by the ODE:

$n$3

where $n$4 is determined by the composition of the satisfaction probabilities and the reinforcement/aggregation function $n$5 (or, for varying $n$6, by its limiting counterpart $n$7). The basin of attraction and limiting set structure are characterized by the invariant sets of this ODE.

Unique Limit and Explicit Criteria

• If $n$8 (or $n$9) is contractive in a suitable sense (expressed as a Lipschitz-type bound weighted by satisfaction parameters), there is a unique deterministic fixed point $A$0 in the state simplex $A$1, to which the process converges almost surely. The unique limit is characterized as the solution to the fixed point equation $A$2 or $A$3, depending on the sampling scenario.
• The sequence of customer discontent proportions converges to $A$4, explicit in the limiting satisfaction rates and product preferences.

Sufficient Conditions on Random Memory Law

• For memory sizes distributed as (truncated) discrete uniform, suitable geometric, binomial, or (truncated) Poisson distributions, the paper provides explicit verification that inverse moments decay sufficiently fast to guarantee almost sure convergence. For example, for $A$5 with $A$6, the series controlling error terms and stochastic approximation noise converge.

Distributional (Weak) Limits and Fluctuations

• Fluctuation results for the properly scaled process $A$7 depend on the eigenvalues of the drift Jacobian, leading to different regimes:
  • For critical eigenvalue $A$8, a refined scaling $A$9 is required, and the normalized process converges in law to a centered normal with explicit covariance.
  • For $B$0, the scaling is $B$1 and the process exhibits degenerate fluctuations (convergence to a random variable along a specific eigendirection, i.e., non-CLT limits).
  • For $B$2, standard $B$3 scaling leads to Gaussian fluctuations, again with explicit covariance matrices derived from the model's parameters.
• All these conclusions apply for both fixed and varying sample size laws, under appropriate regularity and convergence rate conditions (uniformity of polynomial approximation, bounded higher derivatives, etc.).

Methodological Contributions

• The analysis is primarily via the lens of stochastic approximation (SA) and associated ODE methods, supported by martingale convergence theory and the modern theory of randomly reinforced urns.
• The paper develops and applies precise error bounds for polynomial (Bernstein) approximations of the reinforcement functions under both Lipschitz and $B$4 regularity, controlling uniform approximation error in terms of inverse moments of $B$5.
• The limiting distributions are derived using recent results on SA central limit theorems with random step sizes and perturbations [see, e.g., Zhang (2016)].

Connections to Literature and Generalizations

• This work bridges and extends models in the literature on path-dependent random walks, multi-extraction ERW [cf. Franchini (2025)], and generalized urn schemes, including market share dynamics, multi-agent learning, and reinforced stochastic processes.
• The inclusion of attributed steps (binary satisfaction) introduces an additional feedback channel, representing layered reinforcement beyond simple majority or pattern-following rules.
• The random sample size (memory length) is a significant generalization; previous models considered fixed or deterministic $B$6, and did not analyze random $B$7 with arbitrary law and possible time variation.
• The model can be further extended to include "lazy" steps (inactive customers) or product multiplicity (multiple purchases per step), with the authors indicating feasible pathways for these extensions.

Numerical and Analytical Implications

• The conditions for unique almost sure limiting proportions can be translated into explicit numerical procedures (root-finding for the ODE fixed point) once the satisfaction probabilities, memory law, and reinforcement function are specified.
• The results offer insight into stochastic self-reinforcement and path dependence in social choice, market competition, and related settings. Notably, strong path-dependent effects and eventual dominance can arise solely from random sampling and small initial fluctuations, even when product qualities are balanced.

Speculative Extensions and Theoretical Impact

• These probabilistic frameworks are applicable to AI agent-based models with endogenous learning and memory, multi-arm bandits with peer feedback, and complex adaptive systems exhibiting self-induced lock-in.
• Subsequent work may address more general, possibly time-varying, nonparametric reinforcement functions, satisfaction probabilities dependent on history or external signals, or high-dimensional extensions with more products or attribute types.
• The techniques developed here for rigorous control of stochastic approximation with random step sizes can be adapted to other reinforcement-driven processes in machine learning, econometrics, and network science.

Conclusion

The paper presents a comprehensive analysis of a highly general memory-reinforced stochastic process—an elephant random walk with attributed steps and random-size extractions. Through a combination of stochastic approximation, functional analysis, and reinforced urn theory, it provides both qualitative and quantitative characterizations of convergence regimes, fluctuation behaviors, and explicit limiting distributions. The results are robust under broad classes of reinforcement functions and memory size laws, and the modeling framework has wide applicability across stochastic learning, social dynamics, and market modeling. This work constitutes a significant extension of the ERW paradigm, providing methods and structural insights foundational for further research in path-dependent stochastic systems and their implications for understanding reinforcement-driven dynamics in economics and complex adaptive systems.